A graph-based framework for relation propagation and its application to multi-label learning

Label propagation exploits the structure of the unlabeled documents by propagating the label information of the training documents to the unlabeled documents. The limitation with the existing label propagation approaches is that they can only deal with a single type of objects. We propose a framework, named "relation propagation", that allows for information propagated among multiple types of objects. Empirical studies with multi-label text categorization showed that the proposed algorithm is more effective than several semi-supervised learning algorithms in that it is capable of exploring the correlation among different categories and the structure of unlabeled documents simultaneously.